The hot hand in professional darts - Publikationen an ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
J. R. Statist. Soc. A (2020)
183, Part 2, pp. 565–580
The hot hand in professional darts
Marius Ötting, Roland Langrock and Christian Deutscher
Bielefeld University, Germany
and Vianey Leos-Barajas
North Carolina State University, Raleigh, USA
[Received September 2018. Final revision September 2019]
Summary. We investigate the hot hand hypothesis in professional darts in a nearly ideal setting
with minimal to no interaction between players. Considering almost 1 year of tournament data,
corresponding to 167492 dart throws in total, we use state space models to investigate serial
dependence in throwing performance. In our models, a latent state process serves as a proxy
for a player’s underlying form, and we use auto-regressive processes to model how this process
evolves over time. Our results regarding the persistence of the latent process indicate a weak
hot hand effect, but the evidence is inconclusive.
Keywords: Hidden Markov model; Hot hand; Sports statistics; State space model; Time series
1. Introduction
In sports, the concept of a ‘hot hand’ refers to the idea that athletes may enter a state in which
they experience exceptional success. For example, in basketball, players are commonly referred
to as being ‘in the zone’ or ‘on fire’ when they hit several shots in a row. Although empirical
analyses of the hot hand phenomenon tend to focus on sports because the corresponding data
are relatively easily accessible, the notion of the hot hand in fact applies to much more general
settings in which streaks may occur, including human performance in general (Gilden and
Wilson, 1995a), artistic, cultural and scientific careers (Liu et al., 2018), the performance of
hedge funds (Hendricks et al., 1993; Edwards and Caglayan, 2001; Jagannathan et al., 2010),
enduring rivalries in international relations (Gartzke and Simon, 1999; Colaresi and Thompson,
2002) and even gambling activities, against all odds (Xu and Harvey, 2014). However, when
perceiving such dynamics, people tend to overinterpret streaks of success and failure (Bar-Hillel
and Wagenaar, 1991). This phenomenon has been studied intensively by behavioural economists
and psychologists (see, for example, Tversky and Kahneman (1971, 1974)) and is regarded
as a cognitive illusion that has been considered as a primary example for how humans form
beliefs and expectations (Thaler and Sunstein, 2009; Kahneman and Egan, 2011). Especially
in gambling settings it has been demonstrated that people strongly believe in the ‘streakiness’
of their performances, while at the same time also acting according to the gambler’s fallacy,
such that after a streak of identical outcomes an increase in betting volume against the streak is
observed although an independent and identically distributed random process is generating the
outcome (Croson and Sundali, 2005). Such apparent irrationality underlines the importance
Address for correspondence: Marius Ötting, Department of Business Administration and Economics, and
Department of Psychology and Sports Science, Bielefeld University, Bielefeld 33615, Germany.
E-mail: marius.oetting@uni-bielefeld.de
© 2019 The Authors. Journal of the Royal Statistical Society: Series A (Statistics in Society) 0964–1998/20/183565
published by John Wiley & Sons Ltd on behalf of Royal Statistical Society.
This is an open access article under the terms of the Creative Commons Attribution License, which permits use,
distribution and reproduction in any medium, provided the original work is properly cited.566 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas
of being able to quantify precisely a potential hot hand effect in settings where its existence
is highly disputed, e.g. in professional sports. In general, a profound knowledge regarding the
existence and magnitude of streakiness in performances can aid general decision making (Miller
and Sanjurjo, 2018).
In their seminal paper, Gilovich et al. (1985) analysed basketball free-throw data but found
no support for a hot hand, hence coining the notion of the ‘hot hand fallacy’. The alleged
fallacy has been attributed in particular to a potential memory bias, with notable streaks in
performances being more memorable than outcomes that are perceived as random, but also
to general misconceptions regarding chance, with laypeople expecting randomness to lead to
performances that are more balanced in terms of successes and failures than is actually so.
Since Gilovich et al. (1985), there has been mixed evidence regarding the hot hand in sports,
with some papers claiming to have found indications of a hot hand phenomenon and others
disputing its existence. Bar-Eli et al. (2006) reviewed the literature on the hot hand in sports,
including analyses of data from basketball, baseball, golf, tennis, volleyball and bowling. They
summarized 24 studies, from which only 11 studies provide evidence for a hot hand effect.
Perhaps because of the availability of increasingly large data sets, most of the more recent
studies have found evidence for a hot hand effect (see Raab et al. (2012), Shea (2014), Green
and Zwiebel (2017) and Miller and Sanjurjo (2018)), whereas only some studies dispute its
existence by providing mixed results (see Wetzels et al. (2016) and Elmore and Urbaczewski
(2018)).
Two types of approach have been used to investigate such potential hot hand patterns, namely
(a) analyses of the serial correlation of shot outcomes (see, for example, Gilovich et al. (1985),
Dorsey-Palmateer and Smith (2004) and Miller and Sanjurjo (2014)) and
(b) such that use of a latent variable to describe the form of a player (see, for example, Albert
(1993), Sun (2004), Wetzels et al. (2016) and Green and Zwiebel (2017)), where the hot
hand is understood as serial correlation in shot probabilities.
Although there is no consensus in the literature regarding the definition of the hot hand,
Stone (2012) and Miller and Sanjurjo (2018) showed that direct analyses of the correlation
in outcomes, as per (a) above, may vastly underestimate the correlation in shot probabilities.
For example, a correlation of ρp = 0:4 in shot probabilities can co-occur with a very much
lower correlation of ρr = 0:057 in shot realizations (Stone, 2012; Miller and Sanjurjo, 2014).
In other words, if a genuine hot hand process is driven by auto-correlation in success prob-
abilities (i.e. in players’ forms)—which may very well be so—then this can easily go unde-
tected if the focus lies on the (much weaker) serial correlation of outcomes. Stone (2012) and
Arkes (2013) thus concluded that it is preferable to analyse correlation in shot probabilities,
as per (b) above. Hence, in this paper, we focus on approach (b), which we believe is also
more aligned with the way that terminology related to the hot hand concept (e.g. ‘on fire’
and ‘in the zone’) is commonly applied—as argued by Stone (2012), it seems most natural
to measure players’ form by their time-varying success probabilities, rather than (noisy) shot
outcomes.
In addition to such conceptual issues regarding the representation of the hot hand in the
data-generating process, Miller and Sanjurjo (2018) highlighted a subtle selection bias that may
sneak into analyses of sequential data, which provides a further challenge to the findings of
Gilovich et al. (1985). Aside from mathematical fallacies, which would already seem to explain
many failed attempts to prove the existence of the hot hand, we note that many of the existing
studies considered data, e.g. from baseball or basketball, which we believe are hardly suitable for
analysing streakiness in performances. For example, when analysing hitting streaks of a batterHot Hand in Darts 567
in baseball, other factors such as the performance of the pitcher are also important but difficult
to account for. The same applies to basketball, as there are also several factors affecting the
probability of a player to make a shot, e.g. the position (of a field goal attempt) or the effort of
the defence. In particular, an adjustment of the defensive strategy to stronger focus on a player
during a hot hand streak can conceal a possible hot hand phenomenon (Bocskocsky et al.,
2014).
To overcome these caveats, here we investigate whether there is a hot hand effect in profes-
sional darts: a setting with a high level of standardization of individual throws. In professional
darts, well-trained players repeatedly throw at the dartboard from the exact same position and
effectively without any interaction between competitors, making the course of play highly stan-
dardized. In the existing literature, there are very few contributions that consider darts data, and
almost all of these are restricted to laboratory settings. For example, Van Raalte et al. (1995)
analysed the effect of positive and negative self-talk on throwing performances, considering the
throwing sequences of 60 individuals, each of length 15. The hot hand effect has previously
been investigated by using darts data by Gilden and Wilson (1995b); analysing only 24 throwing
sequences of eight volunteers, they found no evidence for a hot hand effect. Here we consider a
much larger data set, with n = 167492 throws in total, which allows for comprehensive inference
regarding the existence and the magnitude of the hot hand effect. Using state space models,
we evaluate serial dependence in a latent state process, which can be interpreted as a player’s
varying form, in line with approaches of type (b) above.
2. Data
A dartboard is divided into 20 numbered slices (1–20 points) and the centre of the board, the
latter with an outer circle (the single bull; 25 points) and an inner circle (the bull’s-eye; 50 points).
Each of the 20 slices is further divided into three segments, the singles, doubles and triples, with
the last two resulting in twice or triple the slice number being awarded as points. All matches
in our analysis were played by two players in the 501-up format. In this format, both players
start with 501 points and take their turns one after another. Within each turn, a player throws
three darts in quick succession, with the value of the segment hit by each dart being reduced
from the current score. The first player to reach exactly 0 points wins a ‘leg’. The last dart used
to reduce the score to 0 must hit a double or the bull’s-eye (‘double out’). (In some matches,
a player must win a prespecified number of ‘sets’, where each set is played as best of five legs.
Whether the match is played in sets or in legs is not relevant for our analysis.) To win the match,
a player must be the first to win a prespecified number of legs (typically between 7 and 15). If a
player wins the match, he advances to the next round of the tournament.
Data were extracted from http://live.dartsdata.com/, covering all professional
darts tournaments organized by the Professional Darts Corporation between April 2017 and
January 2018. In our analysis, we include all players who played at least 50 legs during the time
period considered. This leads to a total of 8310 legs and 167492 dart throws (from 833 matches
played across 25 tournaments).
The data that are analysed in the paper and the programs that were used to analyse them can
be obtained from
https://rss.onlinelibrary.wiley.com/hub/journal/1467985x/series-
a-datasets.
At the beginning of a leg, players consistently aim at high numbers to reduce their points
quickly. The maximum score in a single throw is 60 as in a triple 20 (T20), which players568 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas
usually aim at, but the data indicate that the triples of the numbers 11–19 (T11–T19), and
the bull’s-eye (bull), are targeted in the initial phase of a leg as well. In fact, Tibshirani et al.
(2011) showed that T20 is not necessarily the best segment to aim at, depending on the preci-
sion of a player’s throws. In addition, when the score before the last throw of a turn is slightly
above or around 180, then with that throw players commonly try to avoid ‘bogey numbers’,
i.e. scores below 170 points which cannot be reduced to 0 within a player’s turn. For exam-
ple, if the score is 182 points before the third dart of a turn, then aiming at T20 but hitting
the single 20 would reduce the score to 162 points, which is a bogey number. Hence, with
182 points left, instead of aiming for T20, players may aim for T12, since, if they fail and
hit the single 12, the score reduces to 170 points which can still be checked out with three
darts during the next turn. Thus, in the initial phase of a leg we regard any throw to land
in the set H = {T11, T12, T13, T14, T15, T16, T17, T18, T19, T20, bull} as success. The corre-
sponding empirical proportions of throws hitting the elements of H are displayed in Table 1.
Since a leg is won once a player reaches exactly 0 points, players do not always target H towards
the end of legs, but rather numbers that make it easier for them to reduce to 0. To retain a high
level of standardization and comparability across throws, we thus truncate our time series data,
excluding throws where the remaining score was less than c = 180 points.
p,l
We consider binary time series {yt }t=1,:::,Tp, l , indicating the throwing success of player p
within his lth leg in the data set, with
p,l 1 if the tth throw lands in H,
yt =
0 otherwise,
where the Tp,l th throw is the last throw of player p in his lth leg with the player’s remaining
score still greater than or equal to c = 180. The final data set then comprises n = 167492 throws
by P = 73 players. To illustrate the structure as well as typical patterns of the data, we display
Gary Anderson’s throwing success histories throughout his first 15 legs in the data:
001 011 011
111 110 0
000 111 101
010 000 101 01
000 110 101
111 000 010 0
110 100 101
100 010 010 00
101 010 000 1
110 100 101
101 101 1
001 011 010 0
000 010 010 11
000 001 000 110
000 111 100
Each row corresponds to one leg—truncated when the score fell below 180—and gaps between
blocks of three successive dart throws indicate a break in Anderson’s play when his opponent
was taking his turn. Next we formulate a model that enables us potentially to reveal any unusual
streakiness in the data, i.e. a possible hot hand effect.Hot Hand in Darts 569
Table 1. Absolute frequencies and
proportions for the different outcomes
of H in our data set
Outcome Frequency Proportion
T11 1 0.00001
T12 3 0.00002
T13 1 0.00001
T14 0 0
T15 6 0.00004
T16 4 0.00002
T17 546 0.003
T18 1709 0.010
T19 9897 0.059
T20 53509 0.319
Bull 108 0.0006
3. Modelling the hot hand in darts
3.1. State space model of the hot hand
We aim at explicitly incorporating any potential hot hand phenomenon in a statistical model for
throwing success. As discussed in Section 1, it seems desirable to focus on potential serial corre-
lation in success probabilities. Conceptually, a corresponding hot hand phenomenon naturally
translates into a latent, serially correlated state process, which for any player considered measures
his varying form. For average values of the state process, we would observe normal throwing
success, whereas, for high (or low) values of the state process, we would observe unusually high
(or low) percentages of successful attempts. Figuratively speaking, the state process serves as a
proxy for the player’s ‘hotness’—alternatively, it can simply be regarded as the player’s varying
form. The magnitude of the serial correlation in the state process then indicates the strength
of any potential hot hand effect. A similar approach was indeed used by Wetzels et al. (2016)
and by Green and Zwiebel (2017), who used discrete state hidden Markov models to measure
the form. Although there is some appeal in a discrete state model formulation, most notably
mathematical convenience and ease of interpretation (with cold versus normal versus hot states),
we doubt that players traverse only finitely many states, and we advocate a continuously varying
underlying state variable instead, thus allowing for gradual changes in a player’s form. Specif-
ically, dropping the superscripts p and l for notational simplicity, we consider models of the
form
yt ∼ Bern.πt /, logit.πt / = ηt .st /, st = ht .st−1 / + t , .1/
where {yt }t=1,:::,T is the observed binary sequence indicating throwing success, and {st }t=1,:::,T
is the unobserved continuous valued state process indicating a player’s varying form. We thus
model throwing success by using a logistic regression model in which the predictor ηt .st / for the
success probability πt depends, among other things, on the current form as measured by st . The
unobserved form process {st } is modelled by using an auto-regressive process.
Model (1) is a special case of a state space model, with the dependence structure that is
illustrated in Fig. 1. Under this model, the outcomes of the throws are conditionally independent,
given the player’s form process. However, if there is serial correlation in the form process st , then570 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas
yt−1 yt yt+1 (throwing history)
··· st−1 st st+1 ··· (player’s form)
Fig. 1. Dependence structure of the modelling framework used to investigate a potential hot hand effect:
the throwing history {yt } is an observed binary sequence indicating whether or not individual throws were
successful, whereas the player’s form {st } is an unobserved continuous valued variable which drives the
player’s time-varying success probabilities
this will induce serial correlation also in the throwing performance. This model formulation
is thus in line with the suggestion by Stone (2012) to focus on the auto-correlation of success
probabilities. Crucially, the model includes the possibility to be reduced to the nested special case
of uncorrelated success probabilities, and hence independent observations, corresponding to the
absence of a hot hand phenomenon. Before we discuss how to conduct maximum likelihood
estimation within these types of model, we first present the models that were considered in
the empirical analysis, i.e. the exact forms of ηt .st / and of ht that we use to analyse the darts
data.
3.2. Model specifications
For our hot hand analysis, we formulate three models, which are all special cases of the general
formulation stated in equation (1). We start with formulating a benchmark model (model 1)
which corresponds to the absence of a hot hand effect. To account for differences between the
players, this model includes player-specific intercepts β0,p in the predictor, which indicate the
individual players’ proportions of throwing success (on the logit scale). In principle, a player’s
overall performance level may change over his career but, since our data cover less than 1 year
of tournaments, it is reasonable to assume that this quantity is constant for each player over
the observation period that was considered. Furthermore, we include a categorical covariate
Dt , Dt ∈ {1, 2, 3}, indicating the position of the dart thrown at time t within the player’s current
turn (first, second or third). This categorical covariate is included since the relative frequency
of hitting H, i.e. of throwing success in the early stages of a leg, in fact differs notably across
the three throws within a player’s turn, with the empirical proportions of hitting H in our
data found to be 0.355, 0.409 and 0.420 for the first, second and third throw respectively. The
substantial improvement after the first throw within a player’s turn is partly due to the necessary
recalibration at the start of a turn (but see further discussion below). Hence, the predictor in
model 1, a basic logistic regression model, is given by
logit.πt / = β0,p + β1 I{Dt =2} + β2 I{Dt =3} ,
with I{·} denoting the indicator function, and β0,p player p’s baseline level for the first dart
within any given turn. In model 2, a state variable {st } is included to account for potential
serial correlation in success probabilities (i.e. a hot hand effect). We assume that {st } follows an
auto-regressive process of order 1:
logit.πt / = β0,p + β1 I{Dt =2} + β2 I{Dt =3} + st ,
st = φst−1 + σt ,Hot Hand in Darts 571
with t ∼IID N .0, 1/. Effectively this is a Bernoulli model for throwing success in which the
success probability fluctuates around the players’ baseline levels—β0,p , β0,p + β1 and β0,p + β2
for within-turn throws 1, 2 and 3 respectively—according to the auto-regressive process {st }.
The process {st } can be interpreted as a player’s time-varying form. Thus, at the time of an
individual throw, the success probability of hitting H depends on the general ability of the
player, on the position of the throw within a turn, and on the current (underlying) form as
modelled by the state process. Within this model formulation, a hot hand is defined as auto-
correlation in the state process, i.e. in the current form. For φ = 0, the model reduces to a model
without auto-correlation in the underlying form (i.e. absence of a hot hand), which is similar
to model 1 but includes an additional stochastic component in players’ abilities. Otherwise, if
there is auto-correlation in the underlying form, i.e. if φ > 0, then this would provide evidence
in favour of a hot hand effect in the form of positively correlated success probabilities. For the
beginning of a new leg, we explicitly account for the result of the last leg, by assuming that
s1 ∼ N .μδwon , σδ / if the last leg was won, and s1 ∼ N .μδlost , σδ / if the last leg was lost, i.e. we
assume that a player’s form in a new leg may depend on the result of the last leg, as expressed by
the mean of the initial distribution. It should be noted here that, by pooling the data of different
players, the parameters β1 , β2 , φ, σ, μδwon , μδlost and σδ are assumed to be equal across players,
whereas the ability (as measured by β0,p ) varies between players. In Section 5 we discuss how
these assumptions could be modified.
To consider the structure of a darts match in more detail, we further distinguish between
transitions within a player’s turn to throw three darts (e.g. between the first and second, or
between the second and third throw) and those across the player’s turns (e.g. between the third
and fourth throw). This extension is considered in model 3 and accounts for the fact that there
is a short break in a player’s action between his turns, with the next turn starting with an empty
board, whereas within a single turn the three darts are thrown in very quick succession—it thus
seems plausible that any possible hot hand effect may show time series dynamics within turns
that are different from those across turns. Specifically, within model 3 we assume a periodic
auto-regressive process of order 1 (PAR(1); Franses and Paap (2004)) to describe a player’s
time-varying form:
logit.πt / = β0,p + β1 I{Dt =2} + β2 I{Dt =3} + st ,
φ s + σa t if t mod 3 = 1,
st = a t−1
φw st−1 + σw t otherwise:
Despite distinguishing between transitions within and across turns, this model still allows for
longer-term hot hand effects that last through the whole leg.
Before we present the results of the various models in Section 4, in the next section we first
discuss how to conduct maximum likelihood estimation within the general formulation given
in equation (1).
3.3. Maximum likelihood estimation
The likelihood of a model as in equation (1) involves analytically intractable integration over
the possible realizations of st , t = 1, : : : , T . We use a combination of numerical integration and
recursive computing, as first suggested by Kitagawa (1987), to obtain an arbitrarily fine approxi-
mation of this multiple integral. Specifically, we finely discretize the state space, defining a range
of possible values [b0 , bm ] and splitting this range into m intervals Bi = .bi−1 , bi /, i = 1, : : : , m, of
length .bm − b0 /=m. The likelihood of a single throwing history can then be approximated as572 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas
LT = ... p.y1 , : : : , yT , s1 , : : : , sT /dsT : : : ds1
m
m T
≈ ... Pr.s1 ∈ Bi1 /Pr.y1 |s1 = biÅ1 / Pr.st ∈ Bit |st−1 = biÅt−1 /Pr.yt |st = biÅt /, .2/
i1 =1 iT =1 t=2
with biÅ denoting the midpoint of Bi . This is just one of several possible ways in which the
multiple integral can be approximated (see, for example, Zucchini et al. (2016), chapter 11).
In practice, we simply require that m be sufficiently large. With the specification as a logistic
regression model as in equation (1), we have that
Pr.y |s = bÅ / = [logit−1 {η .bÅ /}]yt [1 − logit−1 {η .bÅ /}]1−yt :
t t it t it t it
The approximate probability of the state process transitioning from interval Bit−1 to interval Bit ,
Pr.st ∈ Bit |st−1 = biÅt−1 /, follows immediately from the specification of ht and the distribution of
the noise, t .
The computational cost of evaluating the right-hand side of equation (2) is of order O.TmT /.
However, the discretization of the state space effectively transforms the state space model into
a hidden Markov model, with a large but finite number of states, such that we can apply the
corresponding efficient machinery. In particular, for this approximating hidden Markov model,
the forward algorithm can be applied to calculate its likelihood at a cost of order O.Tm2 /
only (Zucchini et al. (2016), chapter 11). More specifically, defining δ = .δ1 , : : : , δm / with δi =
Pr.s1 ∈ Bi /, i = 1, : : : , m, the transition probability matrix (TPM) Γ = .γij / with γij = Pr.st ∈
Bj |st−1 = biÅ /, i, j = 1, : : : , m, and m × m diagonal matrix P.yt / with ith diagonal entry equal to
Pr.yt |st = biÅ /, the right-hand side of equation (2) can be calculated as
LT ≈ δP.y1 /ΓP.y2 /: : : ΓP.yT −1 /ΓP.yT /1, .3/
with column vector 1 = .1, : : : , 1/ ∈ Rm . Equation (3) applies to a single leg played by one player.
Assuming independence of the individual leg’s throwing histories, the likelihood of the full data
set is obtained as
Lp
73 p,lp p,lp p,l
L= δP.y1 /ΓP.y2 /: : : ΓP.yTp,pl /1: .4/
p=1 lp =1
We estimate the model parameters by numerically maximizing the approximate likelihood, sub-
ject to the usual technical issues as detailed in Zucchini et al. (2016).
4. Results
For model 1, the benchmark model corresponding to the absence of a hot hand effect, the
estimated player-specific baseline levels for the first within-turn throws, β0,1 , : : : , β0,73 , range
from −1:021 to −0:295, corresponding to throwing success probabilities ranging from 0.265
to 0.427. The coefficients β1 and β2 , which correspond to the increase in throwing success
probabilities after the first throw within a player’s turn (on the logistic scale), are estimated as
0.228 and 0.276 respectively. From the first to the second within-turn throw, there is thus a
strong increase in the chance of success, followed by a further, smaller increase from the second
to the third throw.
Model 2, which unlike model 1 can capture a potential hot hand effect, was fitted by using
m = 150 and −b0 = bm = 2:5 in the likelihood approximation, monitoring the likely ranges of the
process {st } to ensure that the range considered is sufficiently wide given the parameter estimates.Hot Hand in Darts 573
Table 2 displays the parameter estimates (except the player-specific intercepts) including 95%
confidence intervals (CIs) based on the observed Fisher information. Crucially, the estimate
φ̂ = 0:493 seems to support the hot hand hypothesis, indicating considerable serial correlation
in players’ forms, with the associated CI not containing 0. The Akaike information criterion
(AIC) clearly favours the state space formulation, model 2, over the benchmark model assuming
independent throws, model 1 (ΔAIC = 550). However, the state space model as it stands makes
the implicit assumption that observations are regularly sampled, here such that the sampling
unit is one throw. This fails to acknowledge the actual structure of a player’s sequence of throws,
with blocks of three darts being thrown in quick succession, with breaks of a couple of seconds
between blocks (while the opponent takes his turn).
To reflect the grouping of darts better, model 3 distinguishes between transitions within a
player’s turn to throw three darts and those across the player’s turns. For the (approximate)
likelihood of model 3, the TPM Γ is then not constant across time anymore, but equal to either
a within-turn TPM Γ.w/ or an across-turn TPM Γ.a/ . For model 3, which is clearly favoured
over model 2 by the AIC (ΔAIC = 242), the parameter estimates as well as the associated CIs
are displayed in Table 3. The estimate of the persistence parameter of the AR(1) process active
within a player’s turn, φ̂w = 0:727 (95% CI [0:642; 0:811]), corresponds to fairly strong serial
correlation. However, the estimate φ̂a = 0:058 indicates only minimal persistence in the players’
forms across turns. In fact, when at time t a player begins a new set of three darts within a leg,
then the underlying form variable is drawn from an N .0:058st−1 , 0:7892 / distribution, which is
notably close to the two possible initial distributions of the AR(1) process (N .−0:025, 0:6912 /
Table 2. Parameter estimates with 95% CIs
for model 2
Parameter Estimate 95% CI
φ 0.494 [0.438; 0.550]
σ 0.659 [0.565; 0.768]
β1 0.248 [0.221; 0.274]
β2 0.297 [0.269; 0.325]
μδwon −0:051 [−0:102; 0.001]
μδlost −0:068 [−0:117; −0:019]
σδ 0.701 [0.659; 0.745]
Table 3. Parameter estimates with 95% CIs
for model 3
Parameter Estimate 95% CI
φw 0.727 [0.642; 0.811]
φa 0.058 [−0:010; 0.125]
σw 0.461 [0.350; 0.607]
σa 0.789 [0.699; 0.892]
β1 0.270 [0.242; 0.297]
β2 0.331 [0.302; 0.360]
μδwon −0:025 [−0:069; 0.019]
μδlost −0:043 [−0:084; −0:001]
σδ 0.691 [0.648; 0.736]574 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas
and N .−0:043, 0:6912 /), which determine the form at the start of a leg. We cannot rule out that
there may be a weak carry-over effect across turns—our results show no conclusive evidence
in this regard, with the 95% CI for φa only just containing 0. We thus find strong evidence
of serial correlation within turns, whereas the evidence regarding potential carry-over effects
across turns is inconclusive—the relevance of these findings with regard to the hot hand effect
will be discussed in Section 5.
Table 4 provides an overview of the three models fitted, detailing the number of parameters,
the AIC values, the type of state process (if any) and a short description. To check the goodness
of fit and the adequacy of our models, and also to obtain a more detailed picture of the (short-
term) serial correlation that is implied by models 2 and 3, in Table 5 we compare the empirical
relative frequencies of the eight possible throwing success sequences within players’ turns—000,
001, 010, 011, 100, 101, 110 and 111—to the corresponding frequencies as expected under the
three models that were fitted. We restricted this comparison to the first two turns of players
within each leg and used Monte Carlo simulation to obtain the model-based frequencies of the
eight sequences. The benchmark model (model 1), which corresponds to a complete absence of
any hot hand pattern, clearly underestimates the proportion of 000 and 111 sequences, with de-
viations of up to 0.03. This indicates correlation in throwing performances within a turn. Model
2 better reflects the cumulation of 000 and 111 sequences, with a maximum deviation of 0.013.
Finally, model 3, which is favoured by the AIC, almost perfectly captures the proportion of
Table 4. Overview of models 1–3
Model Number of AIC State Description
parameters process
1 75 223200 — Player-specific intercepts and dummy variables
for the throw within a player’s turn
2 79 222651 AR(1) Model 1 + AR(1) state process for the form
3 81 222400 PAR(1) Model 1 + PAR(1) state process, distinguishing
transitions within and across a player’s turn
Table 5. Relative frequencies of the eight possible throwing success
histories within a player’s turn†
Sequence Empirical proportion Results for the following models:
Model 1 Model 2 Model 3
000 0.252 0.222 0.239 0.250
001 0.151 0.159 0.153 0.150
010 0.130 0.152 0.139 0.136
011 0.114 0.111 0.113 0.110
100 0.103 0.121 0.114 0.108
101 0.080 0.088 0.082 0.082
110 0.086 0.084 0.083 0.082
111 0.084 0.063 0.075 0.083
†The second column gives the proportions found in the data, whereas the
last three columns give the proportions as predicted under the various
models fitted, for data structured exactly as the real data.Hot Hand in Darts 575
000 and 111 sequences, with the main mismatch in proportions (0.006) found for 010 sequences.
Whereas Table 5 deals with outcomes within turns, Table 6 provides similar summary statis-
tics corresponding to outcomes across turns. Considering the first two turns by a player within
a leg, Table 6 displays the empirical and model-derived proportions of the 16 possible pairs
.a, b/ ∈ {0, 1, 2, 3}2 corresponding to the number of successes within two consecutive turns; for
example, the pair .1, 3/ indicates that the number of successful throws in the first and second turn
are 1 and 3 respectively. We find that model 3 is clearly superior to model 2 in terms of capturing
the observed proportions of pairs of turns with very different success (e.g. .0, 3/ and .2, 0/). This
is because model 2 assumes that the parameter φ is constant across time, such that the cor-
responding estimate represents a compromise between within-turn and across-turn correlation
found in the data. As a consequence, the former is underestimated, whereas the latter is overesti-
mated, such that the model predicts that pairs of turns with very different success rates will occur
less often than is actually so. In contrast, model 2 overestimates the magnitude of a potential
hot hand effect across turns. In contrast, comparing the empirical proportions of the pairs .0, 0/
and .3, 3/ with the model-derived proportions, we find that model 1, representing the absence of
a hot hand effect, underestimates persistence in success rates not only within turns, as shown in
Table 5, but also across turns. Overall, model 3 thus comes closest to capturing both within-turn
and across-turn correlation in performances. For the observed and expected frequencies that are
shown in Table 6, a corresponding χ2 goodness-of-fit test rejects the null hypothesis at the 1%
level for model 1 (χ2 = 507:1; degrees of freedom df = 15) and for model 2 (χ2 = 130:2; df = 15),
whereas for model 3 the test fails to reject the null hypothesis (χ2 = 29:80; df = 15).
To investigate typical patterns of the hidden process {st } further, we calculate, under model
3, the most likely trajectory of the latent state (i.e. form) for each player and leg. Specifically,
again dropping the superscripts p and l, we seek
Table 6. Relative frequencies of the 16 possible combinations
of numbers of success during the first two turns†
Pair Empirical proportion Results for the following models:
Model 1 Model 2 Model 3
.0, 0/ 0.067 0.051 0.063 0.064
.0, 1/ 0.097 0.096 0.100 0.098
.0, 2/ 0.072 0.062 0.063 0.067
.0, 3/ 0.022 0.013 0.016 0.021
.1, 0/ 0.092 0.096 0.098 0.100
.1, 1/ 0.150 0.186 0.166 0.155
.1, 2/ 0.110 0.122 0.114 0.109
.1, 3/ 0.032 0.027 0.031 0.034
.2, 0/ 0.067 0.062 0.060 0.066
.2, 1/ 0.106 0.122 0.110 0.105
.2, 2/ 0.079 0.082 0.081 0.076
.2, 3/ 0.026 0.018 0.024 0.025
.3, 0/ 0.020 0.013 0.015 0.019
.3, 1/ 0.031 0.027 0.029 0.030
.3, 2/ 0.023 0.018 0.023 0.023
.3, 3/ 0.007 0.004 0.007 0.007
†The second column gives the proportions found in the data, whereas
the last three columns give the proportions as predicted under the
various models fitted, for data structured exactly like the real data.576 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas
0.6
Decoded π *
0.4
0.2
(a)
0.6
Decoded π *
0.4
0.2
(b)
0.6
Decoded π *
0.4
0.2
(c)
0.6
Decoded π *
0.4
0.2
(d)
0.6
Decoded π *
0.4
0.2
(e)
0.6
Decoded π *
0.4
0.2
5 10 15
Throw index t
(f)
Fig. 2. Decoded most likely sequences of throwing success probabilities according to model 3, for more than
100 legs played by each of six players from the data set ( , player-specific intercepts for the respective
player’s within-turn throw 1; , transition between a player’s turn of three darts each): (a) Michael van Gerwen;
(b) Phil Taylor; (c) Gary Anderson; (d) Krzyszt of Ratajski; (e) Zoran Lerchbacher; (f) Peter JacquesHot Hand in Darts 577
.s1Å , : : : , sTÅ / = arg max Pr.s1 , : : : , sT |y1 , : : : , yT /,
s1 ,:::,sT
i.e. the most likely state sequence, given the observations. After discretizing the state space
into m intervals, maximizing this probability is equivalent to finding the optimum of mT pos-
sible state sequences. This can be achieved at computational cost O.Tm2 / by using the Viterbi
algorithm. We then calculate the corresponding trajectories π1Å , : : : , πTÅ of the most likely success
probabilities to have given rise to the observed throwing success histories, taking into account
also the player-specific abilities and the dummy variables. Fig. 2 displays the decoded sequences
for six players from the data set. Since there are only 23 = 8 different possible sequences of ob-
servations within a player’s turn, and since players start each turn almost unaffected by previous
performances (compare φ̂a = 0:058), there is only limited variation in the most likely sequences.
The actual sequences may of course differ from these most likely sequences. The player-specific
intercepts for within-turn throw 1 measure the difference in the players’ abilities; in Fig. 2, the
corresponding success probabilities range from 0.283 (Zoran Lerchbacher) to 0.420 (Michael
van Gerwen). The probability of hitting H increases after the first throw within a turn because
of the two dummy variables. We also see confirmed that the form is not retained across turns.
5. Discussion
Our results indicate that within a player’s turn, involving three darts thrown in quick succession,
there is strong correlation in the latent state process. However, short breaks, which in the given
setting result from the opponent taking his turn, effectively result in a fresh start of the process
describing the player’s form. From a purely statistical point of view, if the hot hand phenomenon
is understood as the presence of serial correlation in individuals’ forms, then our findings would
seem to provide strong evidence in favour of the hot hand. However, some strategic aspects in
darts need to be considered when interpreting our results with regard to a potential hot hand
effect. In particular, depending on the exact position of a turn’s first dart within or close to a
triple segment, this dart can potentially be used as a ‘marker’. Darts 2 and 3 within the same turn
can then be aimed at the marker and may be deflected into the target—the marker thus effectively
increases the target area. The coefficients β̂ 1 and β̂ 2 , which indicate systematic differences in the
success probabilities of the second and third throw within a turn, relative to the first throw, in
fact indicate that the success probability for hitting H increases during a player’s turn (see Tables
2 and 3). However, the corresponding dummy variables do not incorporate any information on
whether or not the first dart is a marker dart—if it is, then the success probability may increase
even further than indicated by β̂ 1 and β̂ 2 . Similarly, the first dart may also end up blocking a
target area, thus decreasing the success probability of subsequent throws. Within our modelling
approach, these effects seem to have been captured by the within-turn serial correlation that
is induced by the latent auto-regressive process. In other words, although we conceptualized
the latent state process as a proxy variable measuring a player’s time-varying form, it seems
likely that within turns this process actually accommodates other not directly observed effects,
namely such as related to marker or blocker darts. The strong serial correlation within a player’s
turn, hence, may be partially tied to the effect of marker darts rather than a hot hand effect.
Since our data do not contain corresponding information that would allow us to disentangle
these two possible causes, there is no definitive conclusion whether the strong serial correlation
within turns provides evidence of a hot hand effect, or rather is a consequence of marker darts.
In any case, it is at least questionable whether serial correlation within a sequence of only three
darts, thrown in quick succession, is what sports commentators, fans and athletes have in mind578 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas
when referring to the hot hand. In conclusion, although we find strong serial correlation for
within-hand throws, the persistence parameter measuring potential correlation across turns was
estimated to be very small, such that overall we do not find conclusive evidence for a hot hand.
Further research specifically on the hot hand in darts could focus on explicitly addressing
player heterogeneity. In addition to the baseline level of πt , the parameters φw , φa , σw and σa ,
and hence the magnitude of the hot hand effect, may vary across players. This could reveal
that for some players the hot hand effect lasts longer than for others, and potentially also
across turns. Modelling this individual variability could be achieved by using covariates or, if
no suitable covariates are available to explain the heterogeneity, via random effects. However,
fitting state space models such as those presented here already involves a high computational
cost, with the numerical maximization taking several days on a usual desktop computer. This
would be further increased when incorporating random effects due to the required integration
over possible values of the random effects. Possible parameter estimation approaches moving
forward with random effects include
(a) using the Laplace approximation to evaluate the marginal likelihood via the template
model builder package in R,
(b) an approximate Bayesian inference approach using an integrated nested Laplace approx-
imation or
(c) using Markov chain Monte Carlo sampling in a Bayesian framework, where the realiza-
tions of the random effects are sampled alongside the other model parameters.
Although the focus of our analyses was on the hot hand hypothesis, darts in fact provide an
excellent setting for studying several other performance-related hypotheses, because of the highly
standardized actions and the absence of interactions between opponents. For example, darts
data have recently been used to investigate how individuals perform in high-pressure situations
(see Deutscher et al. (2018) and Klein Teeselink et al. (2018)).
The modelling framework that was developed in the present paper, with a continuous valued
latent process representing a player’s time-varying form, can easily be tailored to other sports—
or in fact any sequential performance measures—for further investigations into the existence and
magnitude of the hot hand. A caveat of the existing study is the binary nature of the observations,
which corresponds to a rather noisy measure of the actual form. In sports such as archery
and shooting, performance can be measured more precisely by considering the continuous
valued distance between a shot and the middle of the target. Although these sports easily
lend themselves to a time series analysis because of the structured way in which actions take
place, there are many other sports of interest where time intervals between actions are irregular,
including immensely popular sports such as football, American football, rugby or basketball.
With the attention given to these, it would be of interest to transfer our modelling approach
also to corresponding settings with irregular time intervals between actions. Conceptually, this
can relatively easily be achieved by replacing the AR(1) process that was used in this work to
represent the latent form of a player by its continuous time analogue, the Ornstein–Uhlenbeck
process. Via discretizing the state space, the tools that are available for continuous time hidden
Markov models can then be applied for making inference (Jackson et al., 2003). However, given
that evidence from the most recent literature points to a small hot hand effect (see, for example,
Miller and Sanjurjo (2018) and Green and Zwiebel (2017)), we believe that sports settings with
no direct interaction between opponents—e.g. archery, shooting, or free throws in basketball
in an experimental setting as in Gilovich et al. (1985))—are best suited for analyses related to
the hot hand, as otherwise it can be difficult to disentangle hot hand patterns from potential
confounding factors.Hot Hand in Darts 579 Acknowledgements We thank the two reviewers, and especially the Associate Editor and Joint Editor, for their very insightful and constructive comments on earlier versions of this paper. References Albert, J. (1993) A statistical analysis of hitting streaks in baseball: comment. J. Am. Statist. Ass., 88, 1184–1188. Arkes, J. (2013) Misses in “hot hand” research. J. Sprts Econ., 14, 401–410. Bar-Eli, M., Avugos, S. and Raab, M. (2006) Twenty years of “hot hand” research: review and critique. Psychol. Sprt Exer., 7, 525–553. Bar-Hillel, M. and Wagenaar, W. A. (1991) The perception of randomness. Adv. Appl. Math., 12, 428–454. Bocskocsky, A., Ezekowitz, J. and Stein, C. (2014) The hot hand: a new approach to an old “fallacy”. 8th A. Massachusetts Institute of Technology Sloan Sports Analytics Conf. Colaresi, M. P. and Thompson, W. R. (2002) Hot spots or hot hands?: Serial crisis behavior, escalating risks, and rivalry. J. Polit., 64, 1175–1198. Croson, R. and Sundali, J. (2005) The gambler’s fallacy and the hot hand: empirical data from casinos. J. Risk Uncertnty, 30, 195–209. Deutscher, C., Ötting, M., Langrock, R., Gehrmann, S., Schneemann, S. and Scholten, H. (2018) Very highly skilled individuals do not choke under pressure: evidence from professional darts. Preprint arXiv:1809.07659. Bielefeld University, Bielefeld. Dorsey-Palmateer, R. and Smith, G. (2004) Bowlers’ hot hands. Am. Statistn, 58, 38–45. Edwards, F. R. and Caglayan, M. O. (2001) Hedge fund performance and manager skill. J. Fut. Markts, 21, 1003–1028. Elmore, R. and Urbaczewski, A. (2018) Hot and cold hands on the PGA tour: do they exist? J. Sprts Analyt., 4, 275–286. Franses, P. H. and Paap, R. (2004) Periodic Time Series Models. Oxford: Oxford University Press. Gartzke, E. and Simon, M. W. (1999) “Hot hand”: a critical analysis of enduring rivalries. J. Polit., 61, 777–798. Gilden, D. L. and Wilson, S. G. (1995a) On the nature of streaks in signal detection. Cogn. Psychol., 28, 17–64. Gilden, D. L. and Wilson, S. G. (1995b) Streaks in skilled performance. Psychon. Bull. Rev., 2, 260–265. Gilovich, T., Vallone, R. and Tversky, A. (1985) The hot hand in basketball: on the misperception of random sequences. Cogn. Psychol., 17, 295–314. Green, B. and Zwiebel, J. (2017) The hot-hand fallacy: cognitive mistakes or equilibrium adjustments?: Evidence from Major League Baseball. Mangment Sci., 64, 5315–5348. Hendricks, D., Patel, J. and Zeckhauser, R. (1993) Hot hands in mutual funds: short-run persistence of relative performance, 1974–1988. J. Finan., 48, 93–130. Jackson, C. H., Sharples, L. D., Thompson, S. G., Duffy, S. W. and Couto, E. (2003) Multistate Markov models for disease progression with classification error. Statistician, 52, 193–209. Jagannathan, R., Malakhov, A. and Novikov, D. (2010) Do hot hands exist among hedge fund managers?: An empirical evaluation. J. Finan., 65, 217–255. Kahneman, D. and Egan, P. (2011) Thinking, Fast and Slow. New York: Farrar, Straus and Giroux. Kitagawa, G. (1987) Non-Gaussian state-space modeling of nonstationary time series. J. Am. Statist. Ass., 82, 1032–1063. Klein Teeselink, B., Potter van Loon, R., van den Assem, M. J. and van Dolder, D. (2018) Incentives, performance and choking in darts. VU Amsterdam, Amsterdam. (Available from SSRN 3304092.) Liu, L., Wang, Y., Sinatra, R., Giles, C. L., Song, C. and Wang, D. (2018) Hot streaks in artistic, cultural, and scientific careers. Nature, 559, 396–399. Miller, J. B. and Sanjurjo, A. (2014) A cold shower for the hot hand fallacy. Working Paper 518. Innocenzo Gasparini Institute for Economic Research, Bocconi University, Milan. Miller, J. B. and Sanjurjo, A. (2018) Surprised by the hot hand fallacy?: A truth in the law of small numbers. Econometrica, 86, 2019–2047. Raab, M., Gula, B. and Gigerenzer, G. (2012) The hot hand exists in volleyball and is used for allocation decisions. J. Exptl Psychol. Appl., 18, 81–89. Shea, S. (2014) In support of a hot hand in professional basketball and baseball. PsyCh J., 3, 159–164. Stone, D. F. (2012) Measurement error and the hot hand. Am. Statistn, 66, 61–66. Sun, Y. (2004) Detecting the hot hand: an alternative model. In Proc. A. Meet. Cognitive Science Society, vol. 26. Thaler, R. H. and Sunstein, C. R. (2009) Nudge: Improving Decisions about Health, Wealth, and Happiness. New York: Penguin. Tibshirani, R. J., Price, A. and Taylor, J. (2011) A statistician plays darts. J. R. Statist. Soc. A, 174, 213–226. Tversky, A. and Kahneman, D. (1971) Belief in the law of small numbers. Psychol. Bull., 76, 105–110. Tversky, A. and Kahneman, D. (1974) Judgment under uncertainty: heuristics and biases. Science, 185, 1124– 1131.
580 M. Ötting, R. Langrock, C. Deutscher and V. Leos-Barajas Van Raalte, J. L., Brewer, B. W., Lewis, B. P., Linder, D. E., Wildman, G. and Kozimor, J. (1995) Cork!: The effects of positive and negative self-talk on dart throwing performance. J. Sprt Behav., 18, 50–57. Wetzels, R., Tutschkow, D., Dolan, C., van der Sluis, S., Dutilh, G. and Wagenmakers, E.-J. (2016) A Bayesian test for the hot hand phenomenon. J. Math. Psychol., 72, 200–209. Xu, J. and Harvey, N. (2014) Carry on winning: the gamblers’ fallacy creates hot hand effects in online gambling. Cognition, 131, 173–180. Zucchini, W., MacDonald, I. L. and Langrock, R. (2016) Hidden Markov Models for Time Series: an Introduction Using R. Boca Raton: Chapman and Hall–CRC.
You can also read
